alex kerss

@RayFernando1337 Finally, the AI is mature enough to document the problems it confidently invented.

You can teach Composer 2.5 to be a really good QA engineer for your team with this prompt: "go ahead and make a QA section and write that to a doc for phases 0 to 1 with all of these helpful details you have there. That way as we find issues they can be called out with bug reports so that our engineers can fix them. I think we probably want bug reports folder as well so that way qa can write the bug reports to and we can do a bug review and figure out what we need to fix and what can be deferred for engineering."

Hermes has just updated and it looks like they’re really putting effort into solving the issue of agents randomly stopping, getting sidetracked, or just getting stuck. I’ve seen it myself. Agents kind of lose their path and forget what they were trying to do. Although I find Hermes generally needs less babysitting than OpenClaw, it’s still a real issue. This update looks focused on fixing that with stricter automatic resume systems, better heartbeat monitoring, abandoned task recovery, and architecture designed to make goals more persistent. Hopefully this means all my agent lads stop crashing out at random times. Interested to see how much more reliable they become after this update.

I never thought Codex would overtake Claude Code. And I never thought Karpathy would join Anthropic. Both are true today. It’s going to be interesting in these fast moving times what happens next. I think OpenAI’s model has surpassed Anthropic’s but perhaps Karpathy can bring the issue of memory and self improving memory development up to a faster speed. As I think whoever owns the crown in vibe coding in just a few months time is going to need to crack that issue.

Agent memory is a crucial component inside an agentic system. However, I think the current strategy that people are using is leaning too much into the agent harness. For OpenClaw, Hermes agents, Codex, or Claude Code, whatever you're using, I think there has to be some abstraction into a separate application layer where the memory related to that application is stored within the app that the agent is using, not the harness. If the context lives within the app, then other agents can use it too. If it lives in the agent's memory, you've essentially made a private diary, which the organisation of agents may find difficult to tap into.

Agreed, but I’m curious where people think these “data environments” that need to be set up properly actually exist in practice. That seems to be the hard part. Being locked into one system, whatever that system is today, only for better systems to keep emerging as new providers iterate on this problem, feels like a real issue to me.

This is true of all agents, not just coding agents. Probably the biggest challenge that most companies run into in their agent strategy is getting agents the right constrained context to work with for a task. Too much information or conflicting sources, and the agent can easily draw from the data and produce the wrong result. Conflicting sources of truth for documents, data sources that haven’t been kept up to date, knowledge management systems that rely on tribal knowledge to navigate, and so on. On the other end, of course, too little information and the upside is highly limited of agents in the first place. Thus, a lot of challenges with AI strategies are actually data strategy challenges in disguise. This is why there’s such a significant premium on getting structured and unstructured data environments setup properly so agents can work with information effectively. Critical for any large enterprise adopting agents, and also a clear benefit in some cases to startups that can be designed this way from scratch.

@theo Well I got banned, but I would have cancelled my subscription anyway, codex is the new king!

I cancelled my Claude Code sub. I give up.

degen fact If someone livestreams themselves building a SaaS app every day, you can take each transcript incrementally and feed it into ChatGPT. Ask it to continuously update a PRD + implementation spec after every stream. Day by day, stream by stream, you end up with an extremely detailed machine-readable blueprint of the entire app. Feed that into Codex or another coding agent and the reconstruction speed becomes kinda insane. Months of development effort compressed into hours of AI-assisted implementation. Livestreams are accidentally becoming training data for rebuilding software.

I'm building ClawChat, a messaging layer that sits above the agent harness and above applications. ClawChat is about activating agents and connecting those agents to applications. ClawChat is the messaging layer. It feels a bit like WhatsApp but built for AI agents. Not only can you activate AI agents, but you can build up their workspace markdown files from inside ClawChat. There is an applications marketplace where you can connect with one click your agent into an application. I genuinely love building this thing.

@ctantichrist @theo I was using claude -p — that’s why I’m confused. I wasn’t trying to hack around the UI; I was calling the CLI prompt mode from my own tooling.

@jungleskellam @theo Why were you doing this pre this announcemen. Genuine question. You had access to Claude -p.

If you use any of the following with your Claude sub, your usage must got cut by 25x: - T3 Code - Conductor - zed - jean - “Claude -p” in your ci - scripts to call Claude code from other tools They’re disguising this as “free credits”. Don’t fall for it.

She asked me how did you book this appointment I replied my AI. Agent booked it for me a puzzled look appeared on her face. It was a bit uncomfortable, as if she thought I was lying, or I just made up some new set of words. I’m sure in California it would seem okay, but from the town I’m in in the UK, I would suspect 99.9% of the population do not know what’s an AI agent even is

I wanted my YouTube feed to be mostly AI videos. But YouTube keeps feeding you other stuff it knows you’ll click on, so your feed slowly drifts away from what you actually wanted to see. You can try subscribing to loads of AI channels and filtering by subscriptions only, but that gets messy fast. So I built AIYouTubeChannels.com It organizes AI YouTube into 30 different niches so you can browse AI videos by topic, not just by channel. You can filter by views, likes, upload date, and channel size. Click a thumbnail and it takes you straight to YouTube.

@theo I liked the idea I’ve made my agents respond in html, definitely easy to scan responses now. x.com/jungleskellam/…

Is HTML the new Markdown? Had a lot of thoughts on Thariq's latest article so obviously I had to make it a vid

@theo And the solution is…. ?

Security things from the last few days: - CopyFail (linux pwn'd) - CopyFail 2/Dirty Frag - 13 advisories in Next.js - Over 70 CVEs addressed in MacOS 26.5 - ~50 CVEs addressed in iOS 26.5 - YellowKey (Windows Bitlocker pwn'd entirely) - GreenPlasma (Windows privilege escalation) - CVE-2026-21510 and CVE-2026-21513 confirmed to be used by Russia for Windows RCE - CVE-2026-32202 separately confirmed to be used by Russia for sensitive document access - Mini-Shai Hulud (over 300 JS and Python packages compromised via GitHub Action cache poisoning) - Google confirms they have identified AI-powered exploitation of zero days in an unidentified "open-source, web-based system administration too" - Canvas (popular LMS used in most schools) pwn'd entirely - PAN-OS (palo alto networks) pwn'd with a 9.3 severity CVE-2026-0300 Are you scared yet?

Saw @trq212’s post on the unreasonable effectiveness of HTML for agent outputs and immediately implemented it in ClawChat. Agents can now reply with richer HTML cards instead of long markdown walls: clearer plans, capability summaries, onboarding steps, approvals, and workflow state. Markdown for docs. HTML for understanding.

@theo What happens if AI gets better? What happens if AI improves and can write perfect code better than any human? What happens then? I mean all these arguments seem to assume AI will always be AI slop, what happens if AI develops and there is no AI slop?

Coding with agents is a trap, and we all fell for it.

@garrytan Finally made the switch to Hermes, seems to. It have these issues, porting my gang of agent dudes over, Mike Hermes, Jeff Hermes, Mynher Hermes,…. Just a few more dozen then the lads will all be in the Hermes’ harness and rocking away.

I literally went through this today. OpenClaw messed up its own PATH and I spent the morning going into the guts of Dockerfiles and using Claude code to remove the bug... but then I was flying again by afternoon.

Really like the “skillify” idea, and huge respect for the work. My only pushback is that this still feels like a one-agent loop: user asks, agent fails, user notices, skill gets fixed. My chat pattern is worker + auditor + manager agents in the same thread with me, where the auditor catches issues and the manager updates the skill/workflow files, so I get both the corrected answer and the durable fix.

My friend @finbarr says: It's like code as memory. You work with your agent in a non deterministic way to figure out how to execute a task. The first time it does a bunch of research and writes a script and then executes the script. Every future time it faces the same task it just executes the script immediately. The future is already here, it just might not be in your hands unless you decide to build it.

I need an AI agent who can do things for me. Book my daughter’s eye appointment. Contact the doctor about medication. Arrange a plumber to come round. All the little things I end up never doing because I’m disorganised. But the future is not my agent awkwardly phoning a receptionist. And it’s not every local business building some voice AI to answer calls. That’s the wrong model. The real model is simpler. My granddad should be able to set up an agent with one click. Connect it to the local services he uses with one click. Then talk to it like he would talk to his grandson on the phone. The service provider needs the same thing. They don’t need to operate some complicated new app. They just need a messaging interface. A gardener might not even need a dashboard. He might just need his agent, his schedule, and a chat window. Behind the scenes, the agents talk to each other through API endpoints. The consumer talks to their agent in English. The service provider talks to their agent in English. The agents handle the complicated bit. That’s the missing layer. Not voice bots calling humans. Consumer agents. Service provider agents. Frictionless messaging on both sides. Agent-usable applications underneath. That’s where the opportunity is.

@garrytan Interesting read, but my immediate thought was: why not just give an agent access to the PostHog API and talk to the agent in plain English first before rebuilding the whole stack yourselves?

Personal software is coming